Replicate this research

An open invitation to falsify, not to agree.

This project measures Bitcoin's long-lived resource costs and how much of them the fee market internalizes. Its value depends entirely on whether someone else can reproduce the numbers. So: please try to break them.

What to replicate

Quickstart: reproduce the SCCR

git clone https://github.com/prateekposwal/block-space-economics
cd block-space-economics/research/reproduce
python3 reproduce_sccr.py --input input/fee_history_capture.json
bash cross_check.sh          # JS = Python = C? prints the verdict

Expected: avg SCCR 0.289645 (min 0.059036, max 1.294753, 152/155 below 1×), with the three implementations agreeing to < 1e-6. If you do not get that, we want to know — that is the useful result.

The frozen-capture average is 0.240641 × (32000 / 26586) — the whole series

rescales by the N re-base of 2026-09-17 (see the Changelog).

The kit reads N from research/model-spec.json, so a clone reproduces the

current value with no edits.

Reproduce the node population (N)

The canonical N is a measured reachable-node count, not a census — see the population note and verification-population. Reproduce it:

• Reachable count (this is N): the btcnodes reachable-node crawl is the source of record; data/node_census_series.json carries the snapshots and research/model-spec.json → quantities.N the canonical value and date.
• First-party instruments (independent views, each with its own caveat): verification_population.py (the observed/unobservable split), seed_census.py (DNS-seed visible addresses), addrman_churn.py (address-pool persistence), inbound_census.py (an upper bound on private participation — requires the node to serve inbound connections).
• What to check: that every instrument reports a lower bound and states what it does not measure. Addresses ≠ nodes; a reachable crawl is silent about non-listening/private nodes. If a published number implies we enumerated the population, that is a defect — report it.

The dated, hashed state of this thread is pinned in data/population_snapshot.json (rebuilt by tools/research/population_snapshot.py).

What would falsify the claims

• An SCCR materially different from the published value on the same frozen input.
• A node-count N outside ~10K–100K (which would move the level; the direction survives, since SCCR is inverse-linear in N).
• A measured UTXO/state size that contradicts the observed row or the reconstructed table by more than the stated grade allows.
• A fee-market episode in which fees cover storage at a materially higher rate than the published series, sustained.

How to send a result

Open an issue at github.com/prateekposwal/block-space-economics with: the artifact you ran, your inputs, your output, and your environment. Corrections are published in the Research Changelog with their cause — the project's own 10× correction is precedent that this is the intended path, not an embarrassment.

Environmental baseline (what this node ran under)

The inbound census (D5) depends on the node being reachable so peers can dial it back. On the machine that produced this pre-print, that is impossible, and the reason is documented here so a future reproduction audit can compare like with like rather than treating it as a defect.

Both networks available to the host are carrier-grade NAT (CGNAT):

# JioFiber home broadband — four stacked private hops beyond the gateway
1  192.168.29.1          (the local router)
2  10.37.32.1
3  172.16.3.136
4  192.168.161.80
5  192.168.232.132
6  192.168.232.163
# the "public" 49.43.160.42 is the shared CGNAT egress, not a forwardable address

# mobile hotspot — symmetric NAT with multiple carrier egress IPs (STUN)
local 0.0.0.0:58517 -> stun.l.google.com   : 106.67.179.93:50402
                    -> stun.cloudflare.com : 106.67.185.45:50311

Every alternative reachability path was tested and closed: IPv4 port-forward (no forwardable address exists), UPnP / NAT-PMP (router advertises neither), IPv6 inbound (blocked — three independent external probers time out), STUN/ UDP hole-punching (symmetric NAT), free no-card sandboxes (none offers public raw TCP), Play-with-Docker (shut down, HTTP-only), ngrok TCP (card required on free plans), zrok (TCP sharing is private-only).

The inbound census is therefore frozen at 0 (max_concurrent: 0, distinct clearnet IPs: 0) — the correct, unpoisoned value, not a measurement failure. A bare listener on 8333 would receive only port scanners, and counting those as validators would fabricate evidence, so it is explicitly not done.

To reproduce with a distinct-node result you need a host on a non-CGNAT connection. Run python3 tools/net/census.py — it reports DISTINCT (reachable inbound, real un-SNATed peer IPs), CONCURRENCY (Tor/loopback identity is hidden; a lower bound only), or CONTAINED (no inbound path, with the measured reason). Full write-up: research/cgnat-containment-finding.md.

The standard

• Every number is classified Observed / Reconstructed / Modelled with a grade in the Evidence Matrix.
• Methods are in the Methodology paper.
• The checks that guard the data run in tools/research/integrity_audit.py and gate every change.

If a published number cannot be reproduced from the repository, that is a defect in the repository, and we will treat it as one.

← All research · ← Back to Learn